Add Neon implementation of minmax

[Hari Limaye (hazzlim)]

Add and enable a Neon implementation of minmax on ARM64 platforms. The
existing code is refactored to support an unrolled version for
sufficiently large inputs.

This is stacked on #5949.

Performance numbers:

Benchmark	Speedup MSVC	Speedup Clang
bm<uint8_t, Op::Min_val>/8021	26.191	1.073
bm<uint8_t, Op::Min_val>/63	2.833	0.931
bm<uint8_t, Op::Max_val>/8021	24.444	1.04
bm<uint8_t, Op::Max_val>/63	3.209	1.004
bm<uint8_t, Op::Both_val>/8021	66.888	28.622
bm<uint8_t, Op::Both_val>/63	2.219	2.946
bm<uint16_t, Op::Min_val>/8021	13.457	1.023
bm<uint16_t, Op::Min_val>/31	2.25	1.023
bm<uint16_t, Op::Max_val>/8021	13.457	1.023
bm<uint16_t, Op::Max_val>/31	2.588	1.023
bm<uint16_t, Op::Both_val>/8021	38.111	14.178
bm<uint16_t, Op::Both_val>/31	2.324	2.689
bm<uint32_t, Op::Min_val>/8021	23.837	7.554
bm<uint32_t, Op::Min_val>/15	3.896	3
bm<uint32_t, Op::Max_val>/8021	24.318	7.495
bm<uint32_t, Op::Max_val>/15	3.987	2.921
bm<uint32_t, Op::Both_val>/8021	12.5	8.182
bm<uint32_t, Op::Both_val>/15	2.338	2.857
bm<uint64_t, Op::Min_val>/8021	2.01	0.684
bm<uint64_t, Op::Min_val>/7	1.374	1.5
bm<uint64_t, Op::Max_val>/8021	2.115	0.7
bm<uint64_t, Op::Max_val>/7	1.495	1.5
bm<uint64_t, Op::Both_val>/8021	1.029	0.741
bm<uint64_t, Op::Both_val>/7	0.967	1.321
bm<int8_t, Op::Min_val>/8021	26.191	1.073
bm<int8_t, Op::Min_val>/63	3.103	1.111
bm<int8_t, Op::Max_val>/8021	26.111	1.087
bm<int8_t, Op::Max_val>/63	3.286	1.167
bm<int8_t, Op::Both_val>/8021	66.957	28
bm<int8_t, Op::Both_val>/63	1.87	2.574
bm<int16_t, Op::Min_val>/8021	12.866	0.99
bm<int16_t, Op::Min_val>/31	1.952	0.978
bm<int16_t, Op::Max_val>/8021	12.866	1.023
bm<int16_t, Op::Max_val>/31	1.933	0.996
bm<int16_t, Op::Both_val>/8021	28.667	14.178
bm<int16_t, Op::Both_val>/31	2.066	2.42
bm<int32_t, Op::Min_val>/8021	24.405	7.595
bm<int32_t, Op::Min_val>/15	3.983	2.921
bm<int32_t, Op::Max_val>/8021	24.405	7.667
bm<int32_t, Op::Max_val>/15	4.075	2.921
bm<int32_t, Op::Both_val>/8021	14.659	8.182
bm<int32_t, Op::Both_val>/15	2.442	2.921
bm<int64_t, Op::Min_val>/8021	2.16	0.735
bm<int64_t, Op::Min_val>/7	1.6	1.64
bm<int64_t, Op::Max_val>/8021	2.07	0.713
bm<int64_t, Op::Max_val>/7	1.495	1.605
bm<int64_t, Op::Both_val>/8021	1.23	0.761
bm<int64_t, Op::Both_val>/7	0.946	1.35
bm<float, Op::Min_val>/8021	8.928	4.167
bm<float, Op::Min_val>/15	1.971	1.338
bm<float, Op::Max_val>/8021	9.126	4.148
bm<float, Op::Max_val>/15	2.017	1.307
bm<float, Op::Both_val>/8021	4.767	3.855
bm<float, Op::Both_val>/15	0.954	0.933
bm<double, Op::Min_val>/8021	4.563	2.182
bm<double, Op::Min_val>/7	0.854	0.76
bm<double, Op::Max_val>/8021	4.362	2.121
bm<double, Op::Max_val>/7	0.937	0.767
bm<double, Op::Both_val>/8021	2.118	1.964
bm<double, Op::Both_val>/7	0.899	0.47

[Stephan T. Lavavej (StephanTLavavej)]

#5949 has been merged, so this should be ready to be revised.

As a reminder, draft mode disables PR checks when commits are pushed. Merely moving a PR out of draft mode won't trigger checks, so you should mark as "ready for review" and then push commits to trigger PR checks.

[Stephan T. Lavavej (StephanTLavavej)]

I'm mirroring this to the MSVC-internal repo - please notify me if any further changes are pushed.

[Stephan T. Lavavej (StephanTLavavej)]

📉 🦾 📈